Spatio-Temporal Variational Gaussian Processes

Authors: Oliver Hamelijnck, William Wilkinson, Niki Loppi, Arno Solin, Theodoros Damoulas

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We examine the scalability and performance of ST-VGP and its variants. Throughout, we use a Matérn-3/2 kernel and optimise the hyperparameters by maximising the ELBO using Adam [32]." and "Table 1: NYC-CRIME (small) results. ST-SVGP = SVGP when Z is fixed. TRAIN Z MODEL RMSE NLPD ST-SVGP 3.02 0.13 1.72 0.04 SVGP 3.02 0.13 1.72 0.04 ST-SVGP 2.79 0.15 1.64 0.04 SVGP 2.94 0.12 1.65 0.05
Researcher Affiliation Collaboration Oliver Hamelijnck The Alan Turing Institute / University of Warwick ohamelijnck@turing.ac.uk William J. Wilkinson Aalto University william.wilkinson@aalto.fi Niki A. Loppi NVIDIA nloppi@nvidia.com Arno Solin Aalto University arno.solin@aalto.fi Theodoros Damoulas The Alan Turing Institute / University of Warwick tdamoulas@turing.ac.uk
Pseudocode Yes Algorithm 1 Spatio-temporal sparse VGP" and "Algorithm 2 Sparse spatio-temporal smoothing
Open Source Code Yes We provide JAX code for all methods at https://github.com/Aalto ML/spatio-temporal-GPs.
Open Datasets Yes NYC-CRIME Count Dataset We model crime numbers across New York City, USA (NYC), using daily complaint data from [1]. [1] 2014 2015 crimes reported in all 5 boroughs of New York City. https://www.kaggle.com/ adamschroeder/crimes-new-york-city." and "Using hourly data from the London air quality network [29] between January 2019 and April 2019... [29] Imperial College London. Londonair London air quality network (LAQN). https://www.londonair. org.uk, 2020.
Dataset Splits Yes We use 5-fold cross-validation (i.e., 80 20 train-test split), train for 500 iterations (except for AIR-QUALITY where we train for 300) and report RMSE, negative log predictive density (NLPD, see App. K.1) and average per-iteration training times on CPU and GPU.
Hardware Specification No The paper mentions running experiments on 'CPU and GPU' and refers to 'computational resources provided by the Aalto Science-IT project and CSC IT Center for Science, Finland', but does not provide specific hardware details like CPU or GPU models.
Software Dependencies No The paper mentions 'JAX code' but does not specify version numbers for JAX or any other software libraries required for replication.
Experiment Setup Yes We use learning rates of ρ = 0.01, β = 1 in the conjugate case, and ρ = 0.01, β = 0.1 in the non-conjugate case. We train for 500 iterations (except for AIR-QUALITY where we train for 300) and report RMSE, negative log predictive density (NLPD, see App. K.1) and average per-iteration training times on CPU and GPU. ... SVGP with 2000, 2500, 5000, and 8000 inducing points with mini-batch sizes of 600, 800, 2000, and 3000 respectively.